Visual Attention-Based Airport Detection In Remote Sensing Images

Target detection and recognition in remote sensing images becomes a hot topic in the research area of automatic target detection recently. Among these, airport detection attracts lots of attention because of its applications and importance in military and civil aviation fields. Unfortunately, the complicated background around airports brings much difficulty into the detection. Traditional methods on airport detection always need to analyze images pixel by pixel, which is time consuming due to the large size of remote sensing images. However, as we known, human being can analyze and process the information in his eyeshot rapidly and selectively, which is relative to the existence of visual attention mechanism. So this paper aims to introduce the visual attention mechanism into the airport detection and proposes a new scheme on airport detection and recognition which is based on visual attention. Different kinds of image resources are concerned here and the main innovations in this paper can be described as follows:1. According to the characteristics of airports in panchromatic band, we improved the graph-based visual saliency (GBVS) model by adding Hough transform and prior knowledge of intensity, which can extract regions of interest (ROI) better. After that, the scale-invariant feature transform (SIFT) features are extracted on each ROI and hierarchical discriminant regression (HDR) tree is applied as a classifier. Finally the airport is located by the classification result of SIFT features. Experimental results show that the proposed method is faster and more accurate than existing methods, and has lower false alarm rate and better anti-noise performance simultaneously.2. We apply the improved GBVS model to the airport detection in multispectral remote sensing images. The Euclidean distance on the dimension of spectrum is used to define the difference between two pixels in the Markov process, which enables the model to deal with multi-dimensional data. Then we extract SIFT features from the data in panchromatic band. In this way, the problem of airport detection in multispectral remote sensing images can be resolved well.3. Considering the saliency map of an image as a feature descriptor, we combine the saliency feature and the gist feature together to describe the airport region. For a given image, we first segment it into several chips. Then we extract saliency feature and gist feature which are based on the GBVS model from each chip. After combining these two features together, we use the support vector machine (SVM) to classify it. The classification result will decide which chip contains airport. Experimental results show good quality of these features on the description of regions.